基于多序列MR图像融合的脑肿瘤自动分割算法  

Automatic segmentation algorithm for brain tumors based on multi sequence MR image fusion

作　　者：陈梦雨 郭嘉鹏 徐国苏 李敏 朱珊 朱红 CHEN Mengyu;GUO Jiapeng;XU Guosu;LI Min;ZHU Shan;ZHU Hong(School of Medical Information and Engineering,Xuzhou Medical University,Xuzhou 221000,Jiangsu,China)

机构地区：[1]徐州医科大学医学信息与工程学院,江苏徐州221000

出　　处：《智能计算机与应用》2024年第8期121-128,共8页Intelligent Computer and Applications

基　　金：江苏省高等学校大学生创新创业训练计划项目(202010313018Z)。

摘　　要：准确、高效的脑肿瘤分割,对脑肿瘤的精准诊断具有重要意义。由于脑肿瘤MR图像存在对比度低、易出现噪声、偏移场和容积效应等问题,现有脑肿瘤分割模型的分割精度较低。为了提高脑肿瘤分割精度,提出了基于双通道全卷积神经网络和条件随机场的多序列MR图像融合的脑肿瘤分割算法。双通道全卷积神经网络可提取更丰富的图像特征,条件随机场能克服训练过程的局部极小值和输入图片中噪声产生的不利影响。该算法在脑肿瘤分割挑战数据集BRATS2018中测试,其DSC、PPV、Sensitivity系数均较传统分割方法有显著提高。Accurate and efficient brain tumor segmentation is of great significance for the accurate diagnosis of brain tumors.Due to the low contrast,susceptibility to noise,offset field,and volume effects in brain tumor MR images,the segmentation accuracy of existing brain tumor segmentation models is relatively low.In order to improve the accuracy of brain tumor segmentation,a brain tumor segmentation algorithm based on dual channel fully convolutional neural network and conditional random field for multi sequence MR image fusion is proposed.Dual channel fully convolutional neural networks can extract richer image features,and conditional random fields can overcome the adverse effects of local minima during training and noise in input images.This algorithm was tested on the BRATS2018 brain tumor segmentation challenge dataset and showed significant improvements in DSC,PPV,and Sensitivity coefficients compared to traditional segmentation methods.

关 键 词：多序列MR图像融合 脑肿瘤分割 双通道全卷积神经网络 条件随机场 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]